Breaking the Central Bias: Spatially Partitioned Experts for Coordinate-Based Neuroevolution
Romain Claret, Arthur Gygax, Michael O'Neill, Paul Cotofrei, Michael Palma Mendes, Pascal Felber
cs.NE · cs.CV · cs.LG
Quick verdict
Well-controlled: spatial partitioning lifts ES-HyperNEAT MNIST 21%→43% via coverage; mechanism credible, but pooled-stat unit and single dataset temper it.
Summary
This paper diagnoses why ES-HyperNEAT, a coordinate-based indirect encoding, collapses to ~21% mean accuracy on MNIST: evolved networks concentrate on a small central pixel cluster. The authors partition the 784-pixel input into 1–17 contiguous segments, each handled by a separately evolved 'expert' (Independent MoE) versus a single shared network (Shared MoE), and evaluate 14 aggregation strategies. The Independent MoE with 13 experts reaches 43.07% mean accuracy (106% relative gain over the 20.95% baseline), while the Shared MoE reaches only 26.03%, indicating that per-partition independent evolution, not partitioning alone, drives the gain. Validation-free averaging already yields large improvements (35.60% under the SBG condition; 32.93% under Default), so partitioning carries most of the benefit. Receptive-field tracking shows active-pixel coverage expanding from 3.6% (28 pixels) to a 78.6% peak at Ne=11 (71.7% at Ne=13). Controls (same-batch variants) and a validated baseline replication support internal validity; ANOVA and Welch t-tests with large effect sizes are reported. Limitations, largely acknowledged: a single centered dataset that cannot separate coordinate bias from information density; an untested three-mechanism account of the bias; a 20-generation budget that may underlie the claimed 50–60-input capacity ceiling; and absolute accuracy far below gradient-trained baselines. Code and data are released.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Spatially partitioning MNIST inputs into 13 segments handled by independently evolved ES-HyperNEAT experts raises mean accuracy from 20.95% to 43.07% (106% relative gain).
- The gain comes from per-partition independent evolution, not partitioning alone: the Shared MoE with identical partitioning reaches only 26.03%.
- Partitioning rather than aggregation weighting carries most of the gain: validation-free averaging already improves substantially (35.60% under SBG, 70% relative).
- Receptive-field coverage expands from 3.6% (28 pixels) to a 78.6% peak at Ne=11, demonstrating that partitioning eliminates the spatial-concentration collapse.
- Central bias is an architectural limitation of coordinate-based indirect encodings rather than a failure of evolutionary search, motivating a coverage diagnostic and a partitioning remedy.
Strengths
Clean causal dissociation between partitioning and specialization
Independent and Shared MoE partition the input identically, yet only the Independent variant captures the gain (43.07% vs 26.03%, t=29.19, d=7.54, §5.1), isolating per-partition independent evolution as the active ingredient.
Validated baseline replication and variance controls
The Ne=1 replication is statistically indistinguishable from the 2024 monolithic result (t=−0.68, p=.502, d=−0.17); SBG/SBE/SM switches across seven conditions (§4.4) separate fitness-noise from architectural effects, and Default beating SBG (t=2.29, d=0.59, §6.1) yields a non-obvious finding about noise aiding search.
Mechanistic, structural evidence rather than accuracy-only claims
Honest positioning and disclosed limits
Broad aggregation ablation supports the attribution claim
Weaknesses
Unit of analysis in the statistics is unclear and likely pooled
Abstract-level attribution slippage on the 70% averaging claim
Single centered dataset blocks mechanism isolation
Capacity-ceiling story is budget-relative
Independent-vs-Shared comparison carries an acknowledged confound
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper asks whether ES-HyperNEAT's poor MNIST performance (20.95% mean accuracy over 30 runs in prior work, with active pixels confined to a central cluster) is an optimization artifact or an architectural ceiling. Its intervention is a spatially partitioned ensemble inspired by Mixture-of-Experts: the 784-pixel input is split into 1–17 contiguous segments, each handled by a separately evolved expert (Independent MoE) or by one sequentially applied network (Shared MoE), combined via 14 aggregation strategies. The headline claims: 13 experts with F1-based Perf-Weighted aggregation reach 43.07% mean accuracy (106% relative gain); the Shared MoE reaches only 26.03%, so independent per-partition evolution — not partitioning alone — drives the gain; validation-free averaging already yields large gains (35.60% SBG, 32.93% Default); and receptive-field coverage expands from 3.6% to a 78.6% peak, locating the mechanism in restored input coverage. The diagnostic framing (explicitly not a competitive classifier) is consistent throughout.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
Why does ES-HyperNEAT perform so poorly on MNIST?
Because the substrate coordinate system is centered on the image, evolution converges on networks that use only a small central cluster of pixels (about 3.6% coverage), reaching roughly 21% mean accuracy over 30 runs — far below what the 784 inputs could support. The paper argues this is an architectural bias of coordinate-based encodings, not a failure of the evolutionary search itself.
What accuracy does the proposed expert-partitioning approach reach on MNIST?
The Independent MoE with 13 spatial experts reaches 43.07% mean accuracy (up to 49% in the best run) using F1-based weighted aggregation — a 106% relative improvement over the 20.95% monolithic baseline. Absolute accuracy remains far below gradient-trained models, which exceed 90%; the paper frames the result as a diagnosis, not a competitive classifier.
Is the improvement due to Mixture-of-Experts gating?
No. Routing is fixed by geometry rather than learned, and even equal-weighted averaging that uses no validation data yields a large gain (35.60% under the same-batch condition, about 70% relative). The authors attribute most of the benefit to the spatial partitioning forcing full-image feature discovery, with weighted aggregation amplifying it further.
Is code available for this paper?
Yes. The paper states that code, configurations, and data to reproduce all experiments are available at github.com/RomainClaret/es-hyperneat-optimization-studies, built on the public PUREPLES and NEAT-Python libraries with standard MNIST.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.